AgentSkillsCN

kb-meeting-protocol

从 tmp/ 目录中的转录稿中生成会议纪要：先验证输入内容（需提供语音识别转录稿），在 notes/meetings/artifacts/ 下生成各类成果文件，于 notes/meetings/ 下创建会议记录（YYYY-MM-DD-slug.md），并借助 kb.py typos 为 notes/work/typos.md 添加拼写修正。可通过 /meeting-protocol、“会议纪要”、“梳理转录稿”等指令触发。

SKILL.md
--- frontmatter
name: kb-meeting-protocol
description: 'Build meeting protocols from transcripts in tmp/: validate inputs (Voice Recognition transcript required), generate artifacts under notes/meetings/artifacts/, create a meeting note under notes/meetings/ (YYYY-MM-DD-slug.md) with a copy-ready Russian protocol block, and update notes/work/typos.md via kb.py typos. Triggered by /meeting-protocol, "протокол встречи", "разобрать транскрипт".'

kb-meeting-protocol

Use this skill when the user asks to generate a meeting protocol/report from transcripts in tmp/ and move artifacts into the knowledge base.

Notes on sources

  • •“Speech2Text” refers to an optional speech-to-text service. Configure access via scripts/speech2text.py (base URL via SPEECH2TEXT_BASE_URL, token via SPEECH2TEXT_TOKEN or ~/.config/speech2text/token).

Steps

  1. •Anchor time:
bash
date '+%Y-%m-%d %H:%M (%A) %Z'
  1. •Inspect input transcripts:
bash
ls -la tmp
  • •If the user specifies a meeting (name/substring), filter by that substring in filenames.
  • •If no meeting specified: process everything in tmp/ that you can confidently group per meeting.
  1. •Validate transcripts (per meeting):
  • •A meeting MUST have a Voice Recognition transcript.
  • •Detect VoiceRec by content header TEXT| TIME| SPEAKER (do not rely only on filename).
  • •If VoiceRec is missing or ambiguous: stop and ask the user what to do.
  • •Kontur Talk transcript is optional; if present, use it to improve speaker names; if absent, keep SPEAKER_XX and mention the limitation.
  1. •Collect unclear terms/typos (do not guess):
  • •Capture term:proposed_fix + 1 exact quote (1–2 lines) + 1 short note “о чём речь”.
  1. •Generate artifacts (per meeting):
  • •Create: notes/meetings/artifacts/YYYY-MM-DD-<slug>/
  • •Preserve raw exports, but add at the top: H1 + date + source/recording link (when available).
  • •Write:
    • •kontur-tolk-transcription.md (raw)
    • •voice-rec-transcription-raw.md (raw)
    • •voice-rec-transcription.md (sorted by start time; derived from raw)
  • •Then move/remove the corresponding source files from tmp/ so tmp/ stays temporary.
  1. •Generate meeting note (per meeting):
  • •Create: notes/meetings/YYYY-MM-DD-<slug>.md
  • •Include links to the artifacts directory and each artifact file.
  • •Include a copy-ready protocol block (Russian) either at the top or bottom, with this exact structure:

Тема.

Коротко о встрече. (2–4 предложения без буллетов)

Обсуждаемые темы. (1–5 абзацев по 2–4 предложения; буллеты допустимы, но в основном текст)

Итоги (1–2 абзаца по 2–4 предложения; буллеты допустимы)

Следующие шаги/Договоренности

  • •Задача
    • •Ответственный
    • •Сроки
  1. •Update typo glossary:
  • •Before adding a typo, check if it already exists (limit output):
bash
.venv/bin/python scripts/kb.py typos --query "TYPO" | tail -10
  • •Add newly confirmed typos/fixes using:
bash
.venv/bin/python scripts/kb.py typos --add "TYPO" "FIX"
  1. •Index high-signal outcomes in MCP memory (optional, if allowed):
  • •Apply kb-memory policy (safety gate + “don’t duplicate KB”).
  • •Ensure a Note:<path> entity exists (entityType Note): mcp__server-memory__search_nodes, then mcp__server-memory__create_entities if missing.
  • •Add a single pointer: observation via mcp__server-memory__add_observations: 1–2 lines of what was decided + the note path.
  • •Optionally link to Project:<...> / Person:<...> via mcp__server-memory__create_relations if participants/projects are clear.
  • •Keep it minimal; never store raw transcript content in MCP memory.
  1. •Output to the user:
  • •By default, do not paste the entire protocol if it is already in the meeting note.
  • •Print created meeting note path(s): notes/meetings/YYYY-MM-DD-<slug>.md.
  • •Print unclear terms/typos with context for confirmation.